@MASTERSTHESIS\{IMM2004-03282,
    author       = "P. Juhl",
    title        = "Vurdering af st{\o}jreduktionsalgoritmer",
    year         = "2004",
    keywords     = "St{\o}jreduktion, h{\o}reapparat, taleforst{\aa}elighed, lyttefors{\o}g",
    school       = "Informatics and Mathematical Modelling, Technical University of Denmark, {DTU}",
    address      = "Richard Petersens Plads, Building 321, {DK-}2800 Kgs. Lyngby",
    type         = "",
    note         = "Supervised by Assoc. Prof. Jan Larsen",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/3282-full.html",
    abstract     = "Jeg har i dette projekt opbygget en model til beregning af taleforst{\aa}elighed for st{\o}jreduktionsalgoritmer (SRA). Den f{\o}rste del af modellen anvender Olofsson og Hagermans[11] metode til linear separering af tale og st{\o}jsignal. Til fitting af modellen har jeg brugt resultater fra tidligere lyttefors{\o}g [14]. Modellen er blevet afpr{\o}vet p{\aa} forskellige {SRA,} b{\aa}de {SRA} implementeret i MatlabŪ og h{\o}reapparater med indbygget {SRA}. P{\aa} disse {SRA} har jeg beregnet en v{\ae}rdi for taleforst{\aa}eligheden og ulineariteten af {SRA}. Disse beregninger viste at en {SRA} der brugte {TSVD-}algoritmen (Truncated singular value decomposition) var meget ulinear. Dette bet{\o}d at min model til beregning af taleforst{\aa}elighed ikke virkede ordentligt p{\aa} denne {SRA}.

For at verificere om modellens beregning af taleforst{\aa}elighed er korrekt har jeg udf{\o}rt en r{\ae}kke nye lyttefors{\o}g. Ved disse lyttefors{\o}g har jeg testet 2 forskellige {SRA} fra {HA}. Forskellene p{\aa} de 2 {SRA} stemte overens med resultaterne fra min model. Resultaterne fra lyttetesten er sammenlignet med {STI} (Speech Transmission Index) og {PESQ} (Perceptual Evaluation of Speech Quality). Beregningen af taleforst{\aa}elighed med disse modeller var markant forskellig med resultaterne fra mit lyttefors{\o}g.

In English:

In this project, an objective model to predict speech intelligibility (SI) for noise reduction (NR) is described. The first part of the model uses Olofsson and Hagermans[11] method for linear separation of speech and noise. To adjust the model previous result[14] from listening exercises were used. The model is tested with di erent {NR} implemented in hearing aid (HA) and MatlabŪ . Calculation of the signal-to-distortion-ratio (SDR) and {SI} was performed. These calculations showed that the MatlabŪ -implementation of a {NR} using {TSVD} (Truncated Singular Value Decomposition) was very nonlinear. The model could not separate the speech and noise for the {NR} using {TSVD} and the predicted {SI} was not correct.

To verify the models predictions of {SI} listening exercises was performed. These exercises used two different {NR} implemented in {HA}. Statistical test was used to compare the {SI} for the two {NR}. The models predictions of the speech intelligibility for the {NR,} match the result from the exercise. The models prediction was compared to {STI} (Speech Transmission Index) and {PESQ} (Perceptual Evaluation of Speech Quality). The predictions from these methods did not matched the result of the listening exercise."
}